Serena Papi

A Fast Compressed Sensing Approach to 3D MR Image Reconstruction

The problem of high-resolution image volume reconstruction from reduced frequency acquisition sequences has drawn significant attention from the scientific community because of its practical importance in medical diagnosis. To address this issue, several reconstruction strategies have been recently proposed, which aim to recover the missing information either by exploiting the spatio-temporal correlations of the image series, or by imposing suitable constraints on the reconstructed image volume. The main contribution of this paper is to combine both these strategies in a compressed sensing framework by exploiting the gradient sparsity of the image volume. The resulting constrained 3D minimization problem is then solved using a penalized forward-backward splitting approach that leads to a convergent iterative two-step procedure. In the first step, the updating rule accords with the sequential nature of the data acquisitions, in the second step a truly 3D filtering strategy exploits the spatio-temporal correlations of the image sequences. The resulting NFCS-3D algorithm is very general and suitable for several kinds of medical image reconstruction problems. Moreover, it is fast, stable and yields very good reconstructions, even in the case of highly undersampled image sequences. The results of several numerical experiments highlight the optimal performance of the proposed algorithm and confirm that it is competitive with state of the art algorithms.

Nonlinear Filtering for Sparse Signal Recovery From Incomplete Measurements

The problem of recovering sparse signals and sparse gradient signals from a small collection of linear measurements is one that arises naturally in many scientific fields. The recently developed Compressed Sensing Framework states that such problems can be solved by searching for the signal of minimum L 1-norm, or minimum Total Variation, that satisfies the given acquisition constraints. While L 1 optimization algorithms, based on Linear Programming techniques, are highly effective at generating excellent signal reconstructions, their complexity is still too high and renders them impractical for many real applications. In this paper, we propose a novel approach to solve the L 1 optimization problems, based on the use of suitable nonlinear filters widely applied for signal and image denoising. The corresponding algorithm has two main advantages: low computational cost and reconstruction capabilities similar to those of Linear Programming optimization methods. We illustrate the effectiveness of the proposed approach with many numerical examples and comparisons.

Fast Sparse Image Reconstruction Using Adaptive Nonlinear Filtering

Compressed sensing is a new paradigm for signal recovery and sampling. It states that a relatively small number of linear measurements of a sparse signal can contain most of its salient information and that the signal can be exactly reconstructed from these highly incomplete observations. The major challenge in practical applications of compressed sensing consists in providing efficient, stable and fast recovery algorithms which, in a few seconds, evaluate a good approximation of a compressible image from highly incomplete and noisy samples. In this paper, we propose to approach the compressed sensing image recovery problem using adaptive nonlinear filtering strategies in an iterative framework, and we prove the convergence of the resulting two-steps iterative scheme. The results of several numerical experiments confirm that the corresponding algorithm possesses the required properties of efficiency, stability and low computational cost and that its performance is competitive with those of the state of the art algorithms.

Saliency-based keypoint selection for fast object detection and matching

A new approach to rank and select keypoints based on their saliency.Saliency is defined in terms of detectability, repeatability and distinctiveness.Keypoint detector strength and its local descriptor discriminating power are considered.Our experiments show the effectiveness of the proposed approach.An application for real-time pose estimation confirms our results. In this paper we present a new approach to rank and select keypoints based on their saliency for object detection and matching under moderate viewpoint and lighting changes. Saliency is defined in terms of detectability, repeatability and distinctiveness by considering both the keypoint strength (as returned by the detector algorithm) and the associated local descriptor discriminating power. Our experiments prove that selecting a small amount of available keypoints (e.g., 10%) not only boosts efficiency but can also lead to better detection/matching accuracy thus making the proposed method attractive for real-time applications (e.g., augmented reality).

Matrix thresholding for multiwavelet image denoising

Vector thresholding is a recently proposed technique for the denoising of one-dimensional signals by means of multiwavelet shrinkage. It is more suited both to dealing with the multiwavelet vector coefficients and to taking into account the correlations which can be introduced among the starting vector coefficients by the use of a suitable prefilter. Motivated by the successful results of the multiwavelet transform when used in image processing, the aim of this paper is to extend vector thresholding to the two-dimensional case by introducing the notion of matrix thresholding. This new method allows us to easily exploit the “matrix” nature of the two-dimensional multiwavelet transform, and represents the natural extension of vector thresholding to the 2-D case. Afterwards, as the choice of the threshold level is very important in the practical application of thresholding methods, we propose a first attempt to extend the recently introduced method of H-curve to a multiple wavelet setting. The results of extensive numerical simulations confirm the effectiveness of our proposals and encourage us to keep going in this direction with further studies.

Blind cluster structured sparse signal recovery

We consider the problem of recovering a sparse signal when its nonzero coefficients tend to cluster into blocks, whose number, dimension and position are unknown. We refer to this problem as blind cluster structured sparse recovery. For its solution, differently from the existing methods that consider the problem in a statistical context, we propose a deterministic neighborhood based approach characterized by the use both of a nonconvex, nonseparable sparsity inducing function and of a penalized version of the iterative ?1 reweighted method. Despite the high nonconvexity of the approach, a suitable integration of these building elements led to the development of MB-NFCS (Model Based Nonlinear Filtering for Compressed Sensing), an iterative fast, self-adaptive, and efficient algorithm that, without requiring any information on the sparsity pattern, adjusts at each iteration the action of the sparsity inducing function in order to strongly encourage the emerging cluster structure. The effectiveness of the proposed approach is demonstrated by a large set of numerical experiments that show the superior performance of MB-NFCS to the state-of-the-art algorithms. HighlightsWe focus on recovering cluster structured sparse signals from few acquisitions without any information on the signal structure.We consider a compressed sensing approach with a nonconvex nonseparable neighborhood based sparsity inducing function.We solve the corresponding constrained nonconvex minimization problem by integrating an IRl1 scheme into the penalization approach.The convergence of the iterative algorithm to a local minimum of the original problem is guaranteed.The local nature of the approach allows the algorithm to learn the unknown signal structure during the reconstruction process.

A NOTE ON A MATRIX APPROACH TO MULTIWAVELET APPLICATIONS

In recent years, many papers have been devoted to the topic of balanced multiwavelets, namely, multiwavelet bases which are especially designed to avoid the prefiltering step in the implementation of the multiwavelet transform. In this work, we give a simple algebraic proof of how scalar wavelets can be reinterpreted as the most natural balanced multiwavelets, which maintain the good properties of the wavelet bases they come from. We then show how these new bases can be successfully used to apply matrix thresholding for the denoising of images corrupted by Gaussian noise. In fact, this new approach discovers a balanced matrix nature in Daubechies bases, hence obtaining better numerical results with respect to those achieved via scalar thresholding. In particular, this reinterpretation of scalar wavelets as balanced multiwavelets allows us to successfully use the thresholding filters, previously introduced in the scalar case, in a matrix setting.

Grocery product detection and recognition

An approach for candidate pre-selection based on corners and color information.A robust approach for object detection and recognition; Bag of Words and Deep Neural Networks are compared.A post-processing step, used to combine multiple detection of the same object.A deep experimental evaluation on the complex Grozi-120 public dataset. Object detection and recognition are challenging computer vision tasks receiving great attention due to the large number of applications. This work focuses on the detection/recognition of products in supermarket shelves; this framework has a number of practical applications such as providing additional product/price information to the user or guiding visually impaired customers during shopping. The automatic creation of planograms (i.e., actual layout of products on shelves) is also useful for commercial analysis and management of large stores.Although in many object detection/recognition contexts it can be assumed that training images are representative of the real operational conditions, in our scenario such assumption is not realistic because the only training images available are acquired in well-controlled conditions. This gap between the training and test data makes the object detection and recognition tasks far more complex and requires very robust techniques. In this paper we prove that good results can be obtained by exploiting color and texture information in a multi-stage process: pre-selection, fine-selection and post processing. For fine-selection we compared a classical Bag of Words technique with a more recent Deep Neural Networks approach and found interesting outcomes. Extensive experiments on datasets of varying complexity are discussed to highlight the main issues characterizing this problem, and to guide toward the practical development of a real application.

On the Generation of Synthetic Fingerprint Alterations

In this paper we propose some techniques to generate synthetic altered fingerprints and prove the utility of the generated datasets for developing, tuning and evaluating algorithms for altered fingerprint detection/matching. Due to the lack of public databases of altered fingerprints the generation tool proposed (and made freely available) can be a valid instrument to boost research on these challenging problems.

Saliency-Based Keypoint Reduction for Augmented-Reality Applications in Smart Cities

In this paper we show that Saliency-based keypoint selection makes natural landmark detection and object recognition quite effective and efficient, thus enabling augmented reality techniques in a plethora of applications in smart city contexts. As a case study we address a tour of a museum where a modern smart device like a tablet or smartphone can be used to recognize paintings, retrieve their pose and graphically overlay useful information.